I want to make some augmentations on the fly while reading TfRecords and I have been using the tf.numpy_function which I then wrap to tf.function but my training is very slow.